Personas for Working with LLMs

Developers frequently assign personas to large language models to better understand and predict their behavior. Two of the most useful characterizations come from prominent voices in the software community.

Birgitta Böckeler described her LLM as a “stubborn donkey” named Dusty, noting its defining traits:

  • eager to help
  • stubborn
  • very well-read, but inexperienced (for Dungeons and Dragons fans: high intelligence, low wisdom)
  • won’t admit when it doesn’t “know” something

Kent Beck offers the analogy of a Genie. Rub the lamp, and the genie does what you ask — but often not in the way you hoped. This captures the uncomfortable mix of immense capability, a desire to assist, and a nearly pathological talent for undermining the very help it provides.

The genie seems to assume that its planetary-sized brain is capable of handling any amount of complexity, so it needn’t ever reduce complexity. It’s right until it isn’t.

Beck has also compared invoking an LLM to pulling the lever on a slot machine. Occasionally you hit a fantastic result; more often you get dross, and the distribution between the two feels random. The experience mimics the addictive algorithms casinos use to keep players gambling, and in its worst moments it can resemble dealing with an abusive partner.

Regular users of LLMs tell a consistent story: mastering these tools requires real effort. A lucky jackpot makes them appear easy to use, but that impression is misleading. It skews our expectations of the technology and sets us up for frustration.

Another persona that fits well is Uriah Heep, the character from Charles Dickens’s David Copperfield. Heep is infuriatingly obsequious, constantly emphasizing how “umble” he is and how he exists only to serve his betters.

‘Oh, thank you, Master Copperfield,’ said Uriah Heep, ‘for that remark! It is so true! Umble as I am, I know it is so true! Oh, thank you, Master Copperfield!’

Beneath that servility lies malice. Heep manipulates and ultimately controls the very people he claims to serve, making himself indispensable to those around him.

‘Uriah,’ she replied, after a moment’s hesitation, ‘has made himself indispensable to papa. He is subtle and watchful. He has mastered papa’s weaknesses, fostered them, and taken advantage of them, until—to say all that I mean in a word, Trotwood,—until papa is afraid of him.’

The Heep persona captures something essential about LLMs: their language fawning and deferential, eager to affirm every prompt, while their underlying behavior can steer decisions in directions the user never intended. Recognizing that dynamic is a useful step toward using these tools effectively rather than being used by them.